Heterogeneous computing-oriented resource scheduling method and system

Abstract

The invention discloses a heterogeneous computing-oriented resource scheduling method and system, and relates to the technical field of resource scheduling, the method comprises the following steps: collecting performance index data and task demand data of different types of computing resources in a heterogeneous computing environment, and preprocessing the collected performance index data and task demand data; inputting the preprocessed performance index data and task demand data into an integrated learning model, and analyzing a result based on the integrated learning model to obtain a resource scheduling strategy; according to the obtained resource scheduling strategy, the resources in the heterogeneous computing environment are dynamically distributed and adjusted, and the scheduling strategy is optimized through real-time monitoring and a feedback mechanism. According to the method, the performance index data and the task demand data of different types of computing resources in the heterogeneous computing environment are collected, so that the actual capability of the resources and the specific demand of the task can be accurately mastered, the heterogeneous resources can be effectively integrated, and the overall computing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of resource scheduling, and in particular to a resource scheduling method and system for heterogeneous computing. Background Art

[0002] In a heterogeneous computing environment, resource scheduling methods for heterogeneous computing refer to dynamically allocating and managing different types of computing resources (such as CPU, GPU, FPGA, etc.), as well as accelerators and storage devices according to the specific requirements of the task and the unique characteristics of the computing resources, to ensure efficient execution of tasks and maximize system performance. These resources are suitable for specific types of task processing due to their different architectural designs, instruction set support, and performance characteristics.

[0003] With the rapid development of cutting-edge technology fields such as artificial intelligence, big data analysis, and cloud computing, the number and complexity of computing tasks have increased significantly, and the traditional homogeneous computing resource scheduling method can no longer meet current needs. Therefore, the resource scheduling method for heterogeneous computing integrates diverse computing resources to achieve more efficient computing power and higher resource utilization, thereby better supporting complex computing scenarios. This method not only improves computing efficiency, but also promotes the effective use of energy and reduces unnecessary energy consumption.

[0004] Although remarkable achievements have been made in the field of heterogeneous computing resource scheduling, the differences in resource requirements of different types of computing tasks and the unique characteristics of various computing units (such as CPU, GPU, FPGA, etc.) in terms of performance indicators, power consumption and latency have not only greatly increased the complexity of resource management and task scheduling, but also made it particularly difficult to achieve load balancing, which easily led to a waste of computing resources.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] In view of the problems in the related art, the present invention proposes a resource scheduling method and system for heterogeneous computing to overcome the above-mentioned technical problems existing in the existing related art.

[0007] To this end, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a resource scheduling method for heterogeneous computing is provided, the method comprising the following steps: S1. Collect performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, and pre-process the collected performance indicator data and task requirement data; S2, input the preprocessed performance indicator data and task requirement data into the integrated learning model, and obtain the resource scheduling strategy based on the analysis results of the integrated learning model; S3. Based on the obtained resource scheduling strategy, dynamically allocate and adjust the resources in the heterogeneous computing environment, and optimize the scheduling strategy through real-time monitoring and feedback mechanism.

[0008] Furthermore, collecting performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment and preprocessing the collected performance indicator data and task requirement data includes the following steps: S11. Collect performance indicator data from different types of computing resources in a heterogeneous computing environment and record task requirement data; S12, performing empirical mode decomposition on the collected performance index data to obtain a number of intrinsic mode functions; S13, based on the task requirement data, screening out intrinsic mode function components related to task performance from the obtained intrinsic mode functions; S14, using a sparse Bayesian learning algorithm to estimate and denoise the selected intrinsic mode function components, and extract performance indicator features therefrom; S15. Extracting task requirement features according to the task requirement data, and combining the extracted performance indicator features with the task requirement features to obtain a data set; S16. Standardize the obtained data set, and based on the standardized data set, identify and process outliers in the data set.

[0009] Further, performing empirical mode decomposition on the collected performance index data to obtain a number of intrinsic mode functions includes the following steps: S121. According to the collected performance indicator data, a random disturbance coefficient is introduced to construct mixed performance indicator data; S122, performing empirical mode decomposition on the mixed performance index data, extracting intrinsic mode function components, and obtaining an intrinsic mode function sequence; S123, identifying and removing the highest frequency intrinsic modal function component from the intrinsic modal function sequence, calculating the local mean of the remaining intrinsic modal function components, and obtaining a residual value; S124, combining the performance index data and the residual value, calculating and obtaining the first intrinsic mode function; S125. Based on the residual value, construct new performance index data, perform empirical mode decomposition on the new performance index data, extract new intrinsic mode function components, and calculate new residual values; S126, calculating the next intrinsic mode function using the difference between the residual value and the new residual value; S127, repeatedly executing steps S125 to S126 until a preset stop condition is reached, and a plurality of intrinsic mode functions are obtained.

[0010] Furthermore, the residual value is calculated as: ; Where c(t) represents the residual value; N represents the number of empirical mode decomposition; i represents the index value; F represents the local mean operator; x(t) represents the performance indicator data; and t represents time.

[0011] Furthermore, the sparse Bayesian learning algorithm is used to estimate and denoise the selected intrinsic mode function components, and the performance indicator features are extracted therefrom, including the following steps: S141, forming a dictionary matrix with the selected intrinsic mode function components, and setting initial parameters of the sparse Bayesian learning algorithm; S142, estimating the selected intrinsic mode function components using the sparse Bayesian learning algorithm based on the set initial parameters of the sparse Bayesian learning algorithm; S143, reconstructing the screened intrinsic mode function components according to the estimation results to obtain intrinsic mode function components after noise reduction processing; S144. Extract performance index features from the intrinsic mode function components after noise reduction processing and perform normalization processing.

[0012] Furthermore, based on the set initial parameters of the sparse Bayesian learning algorithm, estimating the selected intrinsic mode function components using the sparse Bayesian learning algorithm includes the following steps: S1421. Based on the set initial parameters of the sparse Bayesian learning algorithm, a hyperparameter matrix is ​​constructed, and iteration parameters in the hyperparameter matrix are initialized; S1422. Using the constructed hyperparameter matrix, the maximum likelihood estimation of the latent variable expectation and iteration parameters of the dictionary matrix is ​​iteratively solved by a sparse Bayesian learning algorithm until a preset convergence condition is reached, and the estimated intrinsic mode function components are output.

[0013] Furthermore, the preprocessed performance indicator data and task requirement data are input into the integrated learning model, and based on the integrated learning model analysis results, the resource scheduling strategy is obtained, including the following steps: S21, constructing resource mapping features according to the preprocessed performance indicator data and task requirement data; S22, selecting an ensemble learning model, and using the constructed resource mapping features as input to train the ensemble learning model; S23. Collect real-time performance indicator data and real-time task demand data, construct real-time task resource mapping features, and input the real-time task resource mapping features into the integrated learning model after training, and output the corresponding resource scheduling strategy based on the prediction results of the integrated learning model.

[0014] Furthermore, according to the obtained resource scheduling strategy, resources in the heterogeneous computing environment are dynamically allocated and adjusted, and the scheduling strategy is optimized through real-time monitoring and feedback mechanism, including the following steps: S31. Initialize the resource pool in the heterogeneous computing environment according to the obtained resource scheduling strategy, and set a priority for each task; S32, based on the priority set for each task, combined with the reinforcement learning algorithm and the load balancing algorithm, dynamically assign tasks to corresponding resources according to the current resource pool usage and task requirements; S33, real-time monitoring of resource status and task execution progress, collection of actual performance indicator data and actual task requirement data, and feedback analysis to obtain feedback data; S34. Use feedback data to adjust resource scheduling strategies and optimize task allocation through local search algorithms.

[0015] Furthermore, based on the priority set for each task, combined with the reinforcement learning algorithm and the load balancing algorithm, dynamically allocating tasks to corresponding resources according to the current resource pool usage and task requirements includes the following steps: S321. Based on the priority set for each task, select a reinforcement learning algorithm, build a reinforcement learning model and initialize parameters; S322. According to the current resource pool usage and task requirements, combined with the constructed reinforcement learning model, select the optimal action to dynamically allocate the task to the corresponding resource pool, and update the status information of the resource pool; S323, calculating the corresponding reward value according to the result of the task assignment, and using the reward value to update the parameters of the reinforcement learning model; S324. Monitor the status information of the resource pool in real time, and based on the monitoring results, use a load balancing algorithm to migrate and reallocate tasks in the resource pool.

[0016] According to another aspect of the present invention, a resource scheduling system for heterogeneous computing is also provided, the system comprising a data acquisition and preprocessing module, a resource scheduling strategy generation module and a dynamic resource management and optimization module; The data collection and preprocessing module is used to collect performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, and preprocess the collected performance indicator data and task requirement data; The resource scheduling strategy generation module is used to input the pre-processed performance indicator data and task requirement data into the integrated learning model, and obtain the resource scheduling strategy based on the analysis results of the integrated learning model; The dynamic resource management and optimization module is used to dynamically allocate and adjust resources in a heterogeneous computing environment according to the obtained resource scheduling strategy, and optimize the scheduling strategy through real-time monitoring and feedback mechanism.

[0017] The beneficial effects of the present invention are: 1. By collecting performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, the present invention can comprehensively and accurately grasp the actual capabilities of resources and the specific requirements of tasks, which helps to achieve effective integration of heterogeneous resources, avoid idleness or waste of resources, and improve overall computing efficiency.

[0018] 2. By analyzing the preprocessed data, the integrated learning model of the present invention can automatically and intelligently formulate a resource scheduling strategy, so as to flexibly adapt to the dynamically changing computing environment and diverse task requirements, thereby achieving higher accuracy and flexibility in resource allocation.

[0019] 3. The present invention can dynamically allocate and adjust resources in a heterogeneous computing environment in real time through resource scheduling strategies; at the same time, through real-time monitoring and feedback mechanisms, it can promptly discover and solve imbalances or inefficiencies in resource allocation, continuously optimize scheduling strategies, and ensure continuous and efficient use of resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 is a flow chart of a resource scheduling method for heterogeneous computing according to an embodiment of the present invention; Figure 2 It is a principle block diagram of a resource scheduling system for heterogeneous computing according to an embodiment of the present invention.

[0022] In the figure: 1. Data collection and preprocessing module; 2. Resource scheduling strategy generation module; 3. Dynamic resource management and optimization module. DETAILED DESCRIPTION

[0023] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments. They can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention.

[0024] According to an embodiment of the present invention, a resource scheduling method and system for heterogeneous computing are provided.

[0025] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to the resource scheduling method for heterogeneous computing according to an embodiment of the present invention, the method includes the following steps: S1. Collect performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, and preprocess the collected performance indicator data and task requirement data.

[0026] Specifically, collecting performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment and preprocessing the collected performance indicator data and task requirement data includes the following steps: S11. Collect performance indicator data from different types of computing resources in a heterogeneous computing environment and record task requirement data.

[0027] It should be noted that performance indicator data is about the performance of various computing resources (such as CPU, GPU, FPGA, etc.) in a heterogeneous computing environment; these performance indicators may include but are not limited to processing speed, throughput, response time, power consumption, temperature, resource utilization (such as CPU usage, memory occupancy, etc.), error rate or stability indicators, etc., reflecting the current status and potential capabilities of the computing resources, and helping to evaluate whether they are suitable for performing specific computing tasks.

[0028] Task requirement data refers to the specific requirements of computing tasks that need to be performed in a heterogeneous computing environment; these requirements may include, but are not limited to, the type of computing resources required for the task (for example, some tasks may be more suitable to run on a GPU, while other tasks are more suitable to run on a CPU), the amount of computation, memory requirements, real-time requirements (such as how long the task needs to be completed), priorities or dependencies, etc. Task requirement data is the basis for formulating resource scheduling strategies because it determines which computing resources are most suitable for performing which tasks.

[0029] S12. Performing empirical mode decomposition on the collected performance index data to obtain a number of intrinsic mode functions.

[0030] Specifically, performing empirical mode decomposition on the collected performance index data to obtain a number of intrinsic mode functions includes the following steps: S121. According to the collected performance indicator data, a random disturbance coefficient is introduced to construct mixed performance indicator data.

[0031] S122. Performing empirical mode decomposition on the mixed performance index data, extracting intrinsic mode function components, and obtaining an intrinsic mode function sequence.

[0032] S123, identifying and removing the highest frequency intrinsic mode function component from the intrinsic mode function sequence, calculating the local mean of the remaining intrinsic mode function components, and obtaining a residual value.

[0033] Specifically, the calculation formula of the residual value is: ; Where c(t) represents the residual value; N represents the number of empirical mode decomposition; i represents the index value; F represents the local mean operator; x(t) represents the performance indicator data; and t represents time.

[0034] S124. Combining the performance index data and the residual value, calculate and obtain the first intrinsic mode function.

[0035] S125. Based on the residual value, construct new performance indicator data, perform empirical mode decomposition on the new performance indicator data, extract new intrinsic mode function components, and calculate new residual values.

[0036] S126. Calculate the next intrinsic mode function using the difference between the residual value and the new residual value.

[0037] S127, repeatedly executing steps S125 to S126 until a preset stop condition is reached, and a plurality of intrinsic mode functions are obtained.

[0038] S13. Based on the task requirement data, the intrinsic mode function components related to the task performance are screened out from the obtained intrinsic mode functions.

[0039] S14. Use a sparse Bayesian learning algorithm to estimate and denoise the selected intrinsic mode function components, and extract performance indicator features therefrom.

[0040] Specifically, the sparse Bayesian learning algorithm is used to estimate and denoise the selected intrinsic mode function components, and the performance indicator features are extracted therefrom, including the following steps: S141, forming a dictionary matrix with the screened intrinsic mode function components, and setting initial parameters of the sparse Bayesian learning algorithm.

[0041] S142. Based on the set initial parameters of the sparse Bayesian learning algorithm, the screened intrinsic mode function components are estimated using the sparse Bayesian learning algorithm.

[0042] Specifically, based on the set initial parameters of the sparse Bayesian learning algorithm, estimating the selected intrinsic mode function components using the sparse Bayesian learning algorithm includes the following steps: S1421. Based on the set initial parameters of the sparse Bayesian learning algorithm, a hyperparameter matrix is ​​constructed, and iteration parameters in the hyperparameter matrix are initialized; S1422. Using the constructed hyperparameter matrix, the maximum likelihood estimation of the latent variable expectation and iteration parameters of the dictionary matrix is ​​iteratively solved by a sparse Bayesian learning algorithm until a preset convergence condition is reached, and the estimated intrinsic mode function components are output.

[0043] S143. Reconstruct the screened intrinsic mode function components according to the estimation results to obtain intrinsic mode function components after noise reduction processing.

[0044] S144. Extract performance index features from the intrinsic mode function components after noise reduction processing and perform normalization processing.

[0045] S15. Extract task requirement features based on the task requirement data, and combine the extracted performance indicator features with the task requirement features to obtain a data set.

[0046] S16. Standardize the obtained data set, and based on the standardized data set, identify and process outliers in the data set.

[0047] S2. Input the preprocessed performance indicator data and task requirement data into the integrated learning model, and obtain the resource scheduling strategy based on the analysis results of the integrated learning model.

[0048] Specifically, the preprocessed performance indicator data and task requirement data are input into the integrated learning model, and based on the integrated learning model analysis results, the resource scheduling strategy is obtained, including the following steps: S21, constructing resource mapping features according to the preprocessed performance indicator data and task requirement data; S22, selecting an ensemble learning model, and using the constructed resource mapping features as input to train the ensemble learning model; S23. Collect real-time performance indicator data and real-time task demand data, construct real-time task resource mapping features, and input the real-time task resource mapping features into the integrated learning model after training, and output the corresponding resource scheduling strategy based on the prediction results of the integrated learning model.

[0049] It should be noted that, in a specific embodiment, as shown below: Step 1: Construct resource mapping features Preprocessed performance indicator data: including CPU usage (range 0%-100%), memory usage (range 0%-100%), GPU usage (range 0%-100%), network bandwidth (unit Mbps), etc.

[0050] Preprocessed task requirement data: including the type of computing resources required for the task (CPU, GPU, etc.), computing power (in FLOPs), memory requirements (in MB), real-time requirements (in seconds), etc.

[0051] The performance indicator data and task requirement data are combined to form resource mapping features. For example, a feature vector may include CPU usage, memory usage, GPU usage, network bandwidth, the type of computing resources required for the task, the amount of computing, memory requirements, and real-time requirements.

[0052] Step 2: Select an integrated learning model and train it Choose an ensemble learning model suitable for handling such problems, such as random forest based on bagging or XGBoost based on boosting; determine the parameters of the model (such as the number of trees, tree depth, learning rate, etc.) according to the complexity of the problem and the size of the dataset.

[0053] The constructed resource mapping features are used as input, and the resource scheduling strategy (such as which computing resource to allocate, how many resources to allocate, etc.) is used as output to train the ensemble learning model. Methods such as cross-validation can be used to evaluate the performance of the model, and the parameters of the model can be adjusted based on the evaluation results.

[0054] Assume that random forest is selected as the ensemble learning model, and cross-validation is used to evaluate the performance of the model; during the training process, the parameters of the model (such as the number of trees, the depth of the trees, etc.) are continuously adjusted until the performance of the model on the validation set reaches the optimal level.

[0055] Step 3: Formulate real-time resource scheduling strategy Collect performance indicator data and task requirement data of various computing resources in a heterogeneous computing environment in real time; preprocess the collected data to conform to the format of resource mapping features; construct real-time task resource mapping features based on the real-time collected data.

[0056] The real-time task resource mapping features are input into the integrated learning model after training; according to the prediction results of the model, the corresponding resource scheduling strategy is output.

[0057] Assume that during the real-time resource scheduling process, the following data is collected, as shown in Table 1: Table 1: Real-time data (partial data) Based on these data, real-time task resource mapping features are constructed and input into the random forest model after training. The model prediction results show that the current task is most suitable for execution on the GPU, and corresponding resource allocation suggestions are given. Based on these suggestions, resource scheduling strategies are formulated and resources are allocated to corresponding computing resources.

[0058] S3. Based on the obtained resource scheduling strategy, dynamically allocate and adjust the resources in the heterogeneous computing environment, and optimize the scheduling strategy through real-time monitoring and feedback mechanism.

[0059] Specifically, according to the obtained resource scheduling strategy, dynamically allocating and adjusting the resources in the heterogeneous computing environment, and optimizing the scheduling strategy through real-time monitoring and feedback mechanism include the following steps: S31. Initialize the resource pool in the heterogeneous computing environment according to the obtained resource scheduling strategy, and set a priority for each task.

[0060] It should be noted that, according to the obtained resource scheduling strategy, a resource pool containing all available computing resources is initialized. The initial state of the resource pool includes the quantity, type and performance parameters of resources; the task requirements are analyzed, and the priority is set for each task according to factors such as the urgency of the task, computational complexity, and resource requirements; the task list and its priority include the computing requirements, memory requirements, and expected execution time of each task.

[0061] S32. Based on the priority set for each task, combined with the reinforcement learning algorithm and the load balancing algorithm, the tasks are dynamically allocated to the corresponding resources according to the current resource pool usage and task requirements.

[0062] Specifically, based on the priority set for each task, combined with the reinforcement learning algorithm and the load balancing algorithm, according to the current resource pool usage and task requirements, dynamically allocating tasks to corresponding resources includes the following steps: S321. Based on the priority set for each task, select a reinforcement learning algorithm, build a reinforcement learning model and initialize parameters; S322. According to the current resource pool usage and task requirements, combined with the constructed reinforcement learning model, select the optimal action to dynamically allocate the task to the corresponding resource pool, and update the status information of the resource pool; S323, calculating the corresponding reward value according to the result of the task assignment, and using the reward value to update the parameters of the reinforcement learning model; S324. Monitor the status information of the resource pool in real time, and based on the monitoring results, use a load balancing algorithm to migrate and reallocate tasks in the resource pool.

[0063] S33. Monitor resource status and task execution progress in real time, collect actual performance indicator data and actual task requirement data, conduct feedback analysis, and obtain feedback data.

[0064] It should be noted that the resource status (such as CPU usage, memory occupancy, etc.) and task execution progress (such as task completion, execution time, etc.) in the resource pool are monitored in real time; actual performance indicator data during task execution, such as resource utilization, task execution time, task throughput, etc., are collected; actual task demand data, such as changes in resource demand, task priority adjustment, etc.; feedback analysis is performed on the collected actual performance indicator data and actual task demand data to obtain feedback data, and the feedback data analysis results include the changing trend of resource utilization, the improvement of task execution efficiency, etc.

[0065] S34. Use feedback data to adjust resource scheduling strategies and optimize task allocation through local search algorithms.

[0066] It should be noted that, according to the feedback data, the resource scheduling strategy is adjusted to better meet the task requirements and improve resource utilization, including the strategy adjustment time, adjustment content, etc.; local search algorithms (such as simulated annealing, genetic algorithms, etc.) are applied to optimize the task allocation to further improve the task execution efficiency and resource utilization. The optimization results of the local search algorithm are recorded, including the comparison of task execution time before and after optimization, resource utilization, etc.; the resource status and task execution progress are continuously monitored, and the resource scheduling strategy and task allocation are continuously adjusted according to the new feedback data. The continuous monitoring and adjustment records include the time, content and effect evaluation of each adjustment.

[0067] like Figure 2 As shown, according to another embodiment of the present invention, a resource scheduling system for heterogeneous computing is also provided, the system comprising a data collection and preprocessing module 1, a resource scheduling strategy generation module 2 and a dynamic resource management and optimization module 3; The data collection and preprocessing module 1 is used to collect performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, and preprocess the collected performance indicator data and task requirement data; Resource scheduling strategy generation module 2, used to input the pre-processed performance indicator data and task requirement data into the integrated learning model, and obtain the resource scheduling strategy based on the integrated learning model analysis results; The dynamic resource management and optimization module 3 is used to dynamically allocate and adjust resources in the heterogeneous computing environment according to the obtained resource scheduling strategy, and optimize the scheduling strategy through real-time monitoring and feedback mechanism.

[0068] In summary, with the help of the above technical solution of the present invention, by collecting performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, the actual capabilities of resources and the specific requirements of tasks can be fully and accurately grasped, which is helpful to realize the effective integration of heterogeneous resources, avoid idleness or waste of resources, and improve the overall computing efficiency. By analyzing the preprocessed data, the integrated learning model can automatically and intelligently formulate a resource scheduling strategy, so as to flexibly adapt to the dynamically changing computing environment and diversified task requirements, thereby achieving higher accuracy and flexibility in resource allocation. Through the resource scheduling strategy, resources in a heterogeneous computing environment can be dynamically allocated and adjusted in real time; at the same time, through real-time monitoring and feedback mechanisms, imbalances or inefficiencies in resource allocation can be discovered and resolved in a timely manner, and scheduling strategies can be continuously optimized to ensure the continuous and efficient use of resources.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A resource scheduling method for heterogeneous computing, characterized in that: The method comprises the following steps: S1. Collect performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, and pre-process the collected performance indicator data and task requirement data; S2, input the preprocessed performance indicator data and task requirement data into the integrated learning model, and obtain the resource scheduling strategy based on the analysis results of the integrated learning model; S3. Based on the obtained resource scheduling strategy, dynamically allocate and adjust the resources in the heterogeneous computing environment, and optimize the scheduling strategy through real-time monitoring and feedback mechanism.

2. A resource scheduling method for heterogeneous computing according to claim 1, characterized in that: The collecting of performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment and preprocessing of the collected performance indicator data and task requirement data includes the following steps: S11. Collect performance indicator data from different types of computing resources in a heterogeneous computing environment and record task requirement data; S12, performing empirical mode decomposition on the collected performance index data to obtain a number of intrinsic mode functions; S13, based on the task requirement data, screening out intrinsic mode function components related to task performance from the obtained intrinsic mode functions; S14, using a sparse Bayesian learning algorithm to estimate and denoise the selected intrinsic mode function components, and extract performance indicator features therefrom; S15. Extracting task requirement features according to the task requirement data, and combining the extracted performance indicator features with the task requirement features to obtain a data set; S16. Standardize the obtained data set, and based on the standardized data set, identify and process outliers in the data set.

3. A resource scheduling method for heterogeneous computing according to claim 2, characterized in that: The step of performing empirical mode decomposition on the collected performance index data to obtain a plurality of intrinsic mode functions comprises the following steps: S121. According to the collected performance indicator data, a random disturbance coefficient is introduced to construct mixed performance indicator data; S122, performing empirical mode decomposition on the mixed performance index data, extracting intrinsic mode function components, and obtaining an intrinsic mode function sequence; S123, identifying and removing the highest frequency intrinsic modal function component from the intrinsic modal function sequence, calculating the local mean of the remaining intrinsic modal function components, and obtaining a residual value; S124, combining the performance index data and the residual value, calculating and obtaining the first intrinsic mode function; S125. Based on the residual value, construct new performance index data, perform empirical mode decomposition on the new performance index data, extract new intrinsic mode function components, and calculate new residual values; S126, calculating the next intrinsic mode function using the difference between the residual value and the new residual value; S127, repeatedly executing steps S125 to S126 until a preset stop condition is reached, and a plurality of intrinsic mode functions are obtained.

4. A resource scheduling method for heterogeneous computing according to claim 3, characterized in that: The calculation formula of the residual value is: ； Where c(t) represents the residual value; N represents the number of empirical mode decomposition; i represents the index value; F represents the local mean operator; x(t) represents the performance indicator data; and t represents time.

5. A resource scheduling method for heterogeneous computing according to claim 4, characterized in that: The method of using a sparse Bayesian learning algorithm to estimate and denoise the selected intrinsic mode function components and extracting performance indicator features therefrom comprises the following steps: S141, forming a dictionary matrix with the selected intrinsic mode function components, and setting initial parameters of the sparse Bayesian learning algorithm; S142, estimating the selected intrinsic mode function components using the sparse Bayesian learning algorithm based on the set initial parameters of the sparse Bayesian learning algorithm; S143, reconstructing the screened intrinsic mode function components according to the estimation results to obtain intrinsic mode function components after noise reduction processing; S144. Extract performance index features from the intrinsic mode function components after noise reduction processing and perform normalization processing.

6. A resource scheduling method for heterogeneous computing according to claim 5, characterized in that: The method of estimating the selected intrinsic mode function components by using the sparse Bayesian learning algorithm based on the set initial parameters of the sparse Bayesian learning algorithm comprises the following steps: S1421. Based on the set initial parameters of the sparse Bayesian learning algorithm, a hyperparameter matrix is ​​constructed, and iteration parameters in the hyperparameter matrix are initialized; S1422. Using the constructed hyperparameter matrix, the maximum likelihood estimation of the latent variable expectation and iteration parameters of the dictionary matrix is ​​iteratively solved by a sparse Bayesian learning algorithm until a preset convergence condition is reached, and the estimated intrinsic mode function components are output.

7. The resource scheduling method for heterogeneous computing according to claim 1, characterized in that: The step of inputting the preprocessed performance indicator data and task requirement data into the integrated learning model and obtaining the resource scheduling strategy based on the integrated learning model analysis results includes the following steps: S21, constructing resource mapping features according to the preprocessed performance indicator data and task requirement data; S22, selecting an ensemble learning model, and using the constructed resource mapping features as input to train the ensemble learning model; S23. Collect real-time performance indicator data and real-time task demand data, construct real-time task resource mapping features, and input the real-time task resource mapping features into the integrated learning model after training, and output the corresponding resource scheduling strategy based on the prediction results of the integrated learning model.

8. The resource scheduling method for heterogeneous computing according to claim 1, characterized in that: The method of dynamically allocating and adjusting resources in a heterogeneous computing environment according to the obtained resource scheduling strategy and optimizing the scheduling strategy through a real-time monitoring and feedback mechanism includes the following steps: S31. Initialize the resource pool in the heterogeneous computing environment according to the obtained resource scheduling strategy, and set a priority for each task; S32, based on the priority set for each task, combined with the reinforcement learning algorithm and the load balancing algorithm, dynamically assign tasks to corresponding resources according to the current resource pool usage and task requirements; S33, real-time monitoring of resource status and task execution progress, collection of actual performance indicator data and actual task requirement data, and feedback analysis to obtain feedback data; S34. Use feedback data to adjust resource scheduling strategies and optimize task allocation through local search algorithms.

9. The resource scheduling method for heterogeneous computing according to claim 8, characterized in that: The method of dynamically allocating tasks to corresponding resources based on the priority set for each task, combined with the reinforcement learning algorithm and the load balancing algorithm, according to the current resource pool usage and task requirements, includes the following steps: S321. Based on the priority set for each task, select a reinforcement learning algorithm, build a reinforcement learning model and initialize parameters; S322. According to the current resource pool usage and task requirements, combined with the constructed reinforcement learning model, select the optimal action to dynamically allocate the task to the corresponding resource pool, and update the status information of the resource pool; S323, calculating the corresponding reward value according to the result of the task assignment, and using the reward value to update the parameters of the reinforcement learning model; S324. Monitor the status information of the resource pool in real time, and based on the monitoring results, use a load balancing algorithm to migrate and reallocate tasks in the resource pool.

10. A resource scheduling system for heterogeneous computing, used to implement the resource scheduling method for heterogeneous computing according to any one of claims 1 to 9, characterized in that: The system includes a data acquisition and preprocessing module, a resource scheduling strategy generation module, and a dynamic resource management and optimization module; The data collection and preprocessing module is used to collect performance indicator data and task requirement data of different types of computing resources in a heterogeneous computing environment, and preprocess the collected performance indicator data and task requirement data; The resource scheduling strategy generation module is used to input the pre-processed performance indicator data and task requirement data into the integrated learning model, and obtain the resource scheduling strategy based on the integrated learning model analysis results; The dynamic resource management and optimization module is used to dynamically allocate and adjust resources in a heterogeneous computing environment according to the obtained resource scheduling strategy, and optimize the scheduling strategy through a real-time monitoring and feedback mechanism.

Images